Long Short-Term Memory / Gated Recurrent Unit

Recurrent neural networks with learned 'gates' that let them retain longer histories, widely used for time-series data.

LSTM was introduced by Hochreiter and Schmidhuber in 1997 as an improved recurrent neural network (RNN) — a network that reads a sequence one step at a time and stores information in a hidden state. In an ordinary RNN, gradients shrink exponentially over time steps during training, so the network struggles to remember anything from far in the past; LSTM adds a forget gate, an input gate, and an output gate that control what information is dropped, written, and output, which eases the problem. GRU, proposed by Cho and colleagues in 2014, has only an update gate and a reset gate and no output gate, so it has fewer parameters and runs faster, often matching LSTM's performance. Since Transformers became dominant, LSTM and GRU have taken a back seat in language tasks, but they're still common in robotics: legged locomotion policies often use them to infer terrain and body state from a history of proprioception, and some imitation-learning baselines for behavior cloning use RNNs too.

ExampleOpenAI's Dactyl used policy-gradient reinforcement learning to train an LSTM policy that controlled a five-fingered dexterous hand, teaching it to reorient a block in-hand and solve a Rubik's cube.

Also called

LSTM, GRU

Related

Recurrent Neural Network · Transformer · State Space Model · History Encoder · Vanishing / Exploding Gradients · Dactyl

Sources

Wikipedia: Long short-term memory
Wikipedia: Gated recurrent unit